Applications - VI (Sec. IIIA)
Parallelization of “Basic” Complex Systems
Synchronous and Loosely Synchronous problems perform well on large parallel machines as long as the problem is large enough. For a given machine, there is a typical sub-domain size (i.e. the grain size or size of that part of the problem stored on each node) above which one can expect to get good performance. There will be a roughly constant ratio of parallel speedup to Nnode if one scales the problem with fixed sub-domain size and total size proportional to Nnode.
Unfortunately although this assertion is probably true in most important cases, it has proven very difficult to design and implement productive programming environments that allow the user to realize this goal. That is why we need to write this book even though in principle success is guaranteed
Is this agreed philosophy? No more discussion here at moment